How much does an AI agent cost?

Search the question and you will see answers that run from a small monthly subscription to a large custom project. They rarely agree, and a big reason is that they are pricing different things. What you actually pay is four costs added together: access, usage, building and connecting the agent, and the time someone spends keeping it right.

Below, each cost comes with a real vendor list price where one exists, followed by what keeps the bill under control and what tends to go wrong. Every price here is a public list price as shown on September 30, 2026. Prices change, so check the linked page before you put a number in a budget.

Budget for four separate costs

Each of the four costs is priced in its own way. A quote that covers one of them does not tell you what the other three will cost, so put each one in the budget separately.

Use seat fees for staff tools, not customer-facing agents. Access is the first of the four costs in AI agent pricing, and it is usually sold as a seat fee, a fixed monthly price per person. Salesforce lists its Agentforce add-ons for Sales, Service, and Field Service and Operations at $125 per user per month. The same page lists Salesforce Foundations at $0 to get started.

Seats are predictable, which makes them easy to budget. They suit an agent that helps your own staff inside software they already use. They fit badly when the agent works for your customers, because customers do not have seats.

Free access exists, but it is for trying things out. Google's Gemini API has a free tier, and its pricing page notes that content on the free plan is used to improve Google's products, while content on the paid plan is not. Keep real customer data off any free plan until you have read its terms.

Choose usage fees when your volume is steady. Usage is the second cost, and it rises and falls with how much the agent does. On the same Agentforce pricing page, customer-facing conversations are $2 each. Flex Credits cost $500 per 100,000 credits, and an Agentforce action uses 20 credits (a voice action uses 30). That comes to 10 cents per action, the same figure Salesforce shows in its own example.

If an agent runs on an AI model you pay for directly, the bill is counted in tokens. A token is a small piece of text the model reads or writes. Anthropic lists Claude Haiku 4.5 at $1 per million input tokens and $5 per million output tokens. Google lists Gemini 3.8 Flash at $0.75 per million input tokens and $3.75 per million output tokens, and its output price includes the tokens the model uses to think before it answers. Both vendors charge separately for what goes in and what comes out.

Usage pricing is fair when your volume is steady. It becomes a risk when volume can jump, because the bill jumps with it.

Get the building work quoted system by system. Building and connecting the agent is the third cost. When you compare quotes for AI agent development cost, look first at the connections. An agent can only act on what it can reach, so it needs access to the systems you already run: your customer records, calendar, inbox or job software. A per-token or per-conversation price does not include that integration work.

Scope decides this part of the price. Truline's pricing page names four things that decide what a client pays: the systems to integrate, the sensitivity of the data, the number of workflows, and how much training your team wants. You can read more about how Truline scopes and quotes this work.

Whoever you hire, ask the quote to list each system the agent connects to and what happens when one of those systems changes. A connection that breaks after a software update is a hidden cost if nobody has agreed who fixes it.

Budget weekly hours for upkeep, the fourth cost. After launch, someone reads the results, fixes failures, updates the agent's instructions when your prices or policies change, and handles the cases the agent should not decide alone. That time costs money whether you pay a provider for it or give it to your own staff.

Put it in the budget from the first day as hours per week, with a named person. Without an owner, nobody is paid to notice when the agent goes wrong.

Expect the running cost to grow past the demo

A demo usually shows one clean request. Real work adds steps, tools, retries and volume, and under usage pricing each of those is billed.

An agent can read and write far more than a chat does. In Anthropic's own data, agents typically use about 4 times the tokens of a chat interaction, and multi-agent systems about 15 times. Those figures describe Anthropic's systems, not yours, but the pattern is worth planning for. Every extra step the agent takes is another step on the invoice.

Tools add to the count as well. On the Claude API, the descriptions of the tools an agent can use, and the results those tools return, are billed as input tokens. The same pricing page lists web search at $10 per 1,000 searches, on top of the token costs.

Here is a useful sense of scale. Anthropic's pricing page estimates about $37 to handle 10,000 support tickets on Claude Haiku 4.5, at around 3,700 tokens per conversation. That figure is model usage only. It leaves out building, connecting, other tools and the person who checks the work. Your real number depends on how many conversations you have and how long they run.

List prices also change on published dates. Google's pricing page lists Gemini 3.8 Flash at $0.75 input and $3.75 output per million tokens through December 31, 2026, then $1.50 and $7.50 from January 1, 2027. Same model, same work, twice the price. Budget from the date you will be paying, not the date you signed the contract.

Start small and price the finished task

Keeping AI agent cost per month predictable starts with a small scope. Anthropic's engineering advice for building agents recommends "finding the simplest solution possible, and only increasing complexity when needed." It adds that this might mean not building an agent at all. That is good buying advice too.

Pick one task with a clear finish, such as answering a known set of customer questions or turning a web form into a draft quote. Write down how often it happens today and how long it takes. That gives you a real cost to compare against, and a way to tell whether the agent paid for itself.

Then divide the monthly bill by the number of tasks the agent finished correctly. A low price per action is cheap when the work is done right and expensive when your staff redo half of it. That is the number that tells you what a working agent costs you.

Lower the unit price with smaller models and published discounts

Vendors sell several models at different prices. Try the lower-priced one first on your own examples, and move up only when it fails on work that matters. A bigger model on a simple task often means paying for capability you do not use.

Work that does not need an instant answer, such as overnight summaries, can often run as a batch. Anthropic's Batch API gives a 50% discount on both input and output tokens, and Google lists a Batch API with a 50% cost reduction on its paid tier.

Caching helps when the agent reads the same long instructions on every request. On most Claude models, reading from the prompt cache is billed at one tenth of the normal input price.

Cap the spend and see the bill at 10 times today's volume

Anthropic states that Claude API billing is based on actual monthly usage. With usage billing, a busy month or an agent stuck repeating itself shows up on the invoice. Before anything goes live, agree on a monthly cap, an alert well below it, and the person who receives the alert.

A pilot at low volume can look cheap. Ask the vendor or builder to show the monthly cost at your current volume and at 10 times that volume. If they cannot, you do not yet know what you are agreeing to.

Avoid the buying mistakes that inflate the bill

Each of these mistakes can be caught with one question or one decision before you sign.

Do not budget from the demo price. Ask what a demo price covers. If it covers access or a small amount of usage, it leaves out integration, oversight and your real volume. Treat it as the start of the conversation, not the budget.

Match the pricing model to how the work arrives. Paying per seat for an agent that serves customers means paying for seats nobody sits in. Paying per use for work with unpredictable volume means an unpredictable bill. Pick the unit that rises only when the useful work rises.

Add a second agent only when the first one works. Anthropic says multi-agent systems, which use about 15 times the tokens of a chat, only make economic sense for tasks whose value is high enough to pay for the extra performance. Anthropic's advice on building agents also warns that because agents act on their own, they cost more to run and their mistakes can build on each other. Start with one agent on one job and add more only when the first one is working and measured.

Agree how to check the agent before you pay. Agree how you will test the agent before your customers rely on it, and what result counts as a pass. Then tie the final payment for the build to that pass, so you are not paying in full for an agent that has not shown it works.

Name one owner for failures. Anthropic's warning that agent mistakes can build on each other matters after launch too, because each step an agent takes can build on the one before. Name one person who owns the agent once it is live, who reviews failures each week and who can switch it off.

Get the billing terms in writing before you sign

Ask what unit you are billed for: seats, conversations, actions or tokens. Get the unit in writing.

Ask for the monthly bill at 10 times today's volume. You want a number, not a reassurance.

Ask who pays for retries and failed runs. A task that fails and runs again can be billed twice.

Ask whether there is a hard spending cap, where it is set, what happens when it is reached and who is told.

Ask what monthly upkeep includes, such as instruction updates, fixes after software changes and review of the cases the agent should not decide alone.

Ask what you keep if you leave. Your data, the agent's instructions and your test examples should stay yours.